Upload a manuscript page and Nomikos segments it into written lines, drafts a transcription where a compatible HTR model is available, and hands you a browser editor to correct, review, share, publish, and export. Behind that sit the editor, the API, storage, job state, streaming, and inference that runs on a researcher's laptop or in the cloud, all in this repository.
Nomikos is the whole ecosystem, data, models, and the application to use them: we trained transcription models for Byzantine Greek, Armenian, Syriac, and Coptic, and published them with open line-level datasets for Greek, Armenian, and Syriac on Hugging Face.
Why Nomikos
- Keep the expert in the loop. Models draft segments and transcriptions; researchers correct and approve. Model output is never treated as automatic ground truth.
- Operate inside your own data boundary. Keep application data behind an API you control, with your own hosting, data policy, and review conventions.
- Run inference on your own computer. Point the
nomikos agent at the platform and run supported BLLA, Calamari, and PP-OCR models on a researcher's CPU. The agent only makes outbound requests, so it needs no inbound port, VPN, or proxy.
- Collaborate through projects and sharing. Organize work in projects, share documents with colleagues, and publish read-only views behind a secret link that a reader opens without an account.
- Correct rather than retype. Fixing a model draft is faster than transcribing a blank page, and approved work exports in a predictable format for publication or retraining.
Built for Research
Nomikos is being developed for the Nomos research ecosystem, with a focus on Syriac, Coptic, Armenian, Byzantine Greek, and related scripts. The platform itself is script agnostic: if you want to collaborate or develop models for your own script, get in touch via GitHub or the Hugging Face organization.
The system is expert-in-the-loop by design. Models draft, and researchers decide what is correct. Approved work produces processed line images and transcription files for publication or future model training.
Complete Workflow
- Turn pages into editable data. Upload or open a page, segment it into lines, generate a model transcription, and pair text with segments.
- Watch the queue while you work. Edit the draft in the browser editor, with job state that says which host ran each job and what is still queued.
- Review, share, and publish. Move documents through review, share them with collaborators, and publish read-only views for readers.
- Export training-ready data. Produce processed line images and transcription files from approved work for publication or future model training.
- Extend to new scripts. Add models through the registry, weights, and publishing workflow. Data preparation, training, and Hub publishing tools are in the repo.
Current Model Support
Through the pinned runtime registry, Nomikos pages can use:
In the platform pickers the segmenters appear as seg-v1 (blla-segment) and seg-v2 (ppocr-segment); each picker starts on the project's default model, else seg-v1, and there is no Default entry.
All four Calamari HTR checkpoints share one architecture: a 40/60-filter CNN followed by two 200-unit
bidirectional LSTM layers at line height 48. Read off the Coptic and Syriac checkpoints, whose
structural fields agree exactly (3x3 convolutions 1 to 40 to 60, BiLSTM weights at layers.4
and layers.6 with 200-unit hidden states, blank index 0, temperature -1.0); each trains
its own codec.
A model is runtime-supported only after its weights are published, pinned, verified, and added to nomikos_inference/registry.yaml. Public weights live on Hugging Face and are cached under ~/.nomikos/hf/cache on first inference. See models and datasets and the publishing workflow for the pinning, verification, and release steps.
Run supported inference locally:
uv tool install nomikos-inference # or: pip install nomikos-inference
nomikos pair # links this machine to your account
nomikos run # takes pages from the queue until you stop it
Point it at a different platform with NOMIKOS_API_URL or --api-url.
Accuracy
Test character error rate (CER) of each transcription model. Lower is better. Values come from the Hugging Face model cards, except syriac-ppocr-v1, whose approximate figure is reported by the model's authors and is not directly comparable.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
